【作者】 郑克强；

【导师】 王伟；

【作者基本信息】 大连理工大学 ， 控制理论与控制工程， 2009， 硕士

【摘要】 对于工作在非结构化场景中的移动机器人系统,能否对其周围环境进行有效辨识、感知和重构是其能够自主运行的前提条件。尤其对三维非结构化场景的建模与重构的研究,具有更重要的研究意义和广泛的应用价值。本文主要完成了三维场景重构系统的构建,有效建立了场景的几何特征模型和三维栅格模型,实现了对三维场景的重构与回溯等研究工作。本文利用三维激光扫描数据的有序性,快速提取出非结构化场景中的直线、平面以及边缘轮廓等几何特征,建立三维场景的几何特征模型。为了实现大范围场景数据的调度及可视化,基于八叉树空间划分的思想,建立场景的三维栅格模型,将空间中的激光数据划分到三维栅格中。经典ICP算法通过查找最近点构成匹配对,以迭代的方式实现场景匹配。该方法具有迭代速度慢,且易收敛到局部极小值等缺点。本文提出了使用边缘特征点代替原始点云等五点改进策略,有效地克服了经典ICP算法的不足。为解决ICP算法收敛到局部极小值的缺点,需要利用扫描视点位姿进行初始校正。然而,扫描视点位姿往往不能有效提供,且存在较大误差。因此本文提出利用视觉图像处理的方法,辅助三维激光场景的匹配,从而实现无位姿的场景匹配问题。将视觉摄像机与三维激光测距仪结合,通过标定的方法标定出他们之间参数关系,得到视觉图像与三维激光深度图像之问的映射。首先对两幅场景对应的视觉图像进行处理,利用SIFT算法查找两幅图像中的特征匹配对,再将特征匹配对映射到激光场景中,构成三维激光特征匹配对,最后计算出两幅场景的转换关系,实现无位姿场景匹配。基于三维栅格模型,本文解决了大范围场景的数据分块及实时调度显示的问题。根据OpenGL空间中的虚拟摄像机当前位置,按离摄像机的远近分5个显示等级,实时调度并显示栅格模型中相应的数据。通过实验证明,本文方法能够有效实现非结构化场景的建模与重构,具有很强的实用性。更多还原

【Abstract】 Effective recognition and perception of environment are two essential tasks for mobile robot’s navigation and exploration while working in the unstructured scenes. Especially for 3D scenes modeling and reconstruction, the study has a wide range of applications. The intelligent equipment for 3D scene reconstruction is built in this thesis. Moreover, the geometrical model and 3D grid model are built for 3D scenes. This thesis completes the reconstruction and back review of the 3D scenes.According to the orderliness of laser scanning, the geometrical model is built by extracting the lines, planes, edge contours etc. To complete the data dispatching and visualization in large-scale scenes, a 3D grid model is built by using the octree space partition idea, which divides the laser data into small 3D grids effectively.The classical ICP algorithm completes scene matching iteratively by finding the nearest points to constitute matching pairs. This approach has the disadvantages of slow iterating and converges to a local minimum easily. In this thesis we propose five ways to improve these disadvantages. The scanning pose is needed while matching scenes using ICP. Unfortunately, the scanning pose can not be provided effectively in the most time. Hence, a new method is proposed in this thesis to solve this problem. In the research, a camera is used together with the 3D laser scanner. The mapping relationship of the digital image and the laser range image is obtained by calibrating the parameters between the two kinds of sensors. After the matching pairs in the corresponding digital images are found by the SIFT algorithm, we map them to the laser range image to get the 3D laser matching pairs. At last, the transformation of the two scenes is calculated without scanning pose.Based on the 3D grid model, the data blocking and dispatching of large-scale scenes are solved. According to the distance of grid to the virtual camera in OpenGL space, we divide five areas to dispatch the laser data in 3D grid model. The nearer to the virtual camera, the more data is dispatched and displayed on the screen. A series of experiment results and data analysis show the method’s validity and practicability.更多还原